Three correlations can each lie between minus one and one while failing to describe any possible joint uncertainty model. The missing check is collective: every weighted combination of the inputs must have a non-negative variance.
The short answer is to assemble the complete correlation matrix and test whether it is positive semidefinite. For three standardised inputs, that condition yields an explicit allowed range for the third correlation once the first two are fixed. It is a constraint on what the numbers can mean together, rather than another preference for tidy tables.
This guide extends Uncertainty Quantification with original exact examples, constructive checks and changed-context practice. All correlations and uncertainty budgets are fictional. The lesson concerns model coherence; it does not infer a real instrument’s dependence or recommend a particular numerical repair for an actual dataset.
eduKateSG · Uncertainty Quantification
Find your next learning step
Test proposed correlations collectively, construct compatible alternatives and preserve their evidence.
ROUTE 1 · CHAPTERS 1–5
Build the joint constraint
Define the quantity and reconstruct the first worked relationship.
ROUTE 2 · CHAPTERS 6–10
Derive and construct valid models
Follow the original examples and check why the operation is valid.
ROUTE 3 · CHAPTERS 11–16
Inspect provenance and transformations
Change the evidence and identify which assumptions still hold.
ROUTE 4 · CHAPTERS 17–21
Test model choices and limits
Check uncertainty, scope and independent model alternatives.
ROUTE 5 · CHAPTERS 22–26
Practise and continue
Repair common errors, attempt transfer and choose connected reading.
Full chapter contents · Uncertainty Quantification subject guide
Chapter contents
Build the joint constraint · Chapters 1–5
1. Begin with three jointly defined inputs
2. Assemble the complete matrix
3. Derive the non-negative variance requirement
Derive and construct valid models · Chapters 6–10
6. Obtain the three-input determinant condition
7. Solve for the allowable third correlation
8. Explain the bound through residuals
Inspect provenance and transformations · Chapters 11–16
11. Understand the upper boundary
12. Inspect a compact set of alternatives
13. See how pairwise deletion can create a conflict
14. Distinguish a common-data covariance estimate
Test model choices and limits · Chapters 17–21
17. Interpret eigenvalues without losing the meaning
18. Treat numerical repair as a changed model
19. Preserve uncertainty about the correlations themselves
Practise and continue · Chapters 22–26
22. Handle a perfect first correlation without dividing by zero
23. Distinguish a valid pooled model from the current regime
24. Notice when two supplied correlations leave the third unresolved
CHAPTER 1 OF 26 · Build the joint constraint
1. Begin with three jointly defined inputs
Let X, Y and Z be uncertain quantities with finite, positive variances. For the first examples, subtract their respective means and divide by their standard deviations. Each resulting standardised variable has mean zero and variance one. We continue using the short labels X, Y and Z for those standardised quantities.
Standardisation makes the correlations equal to the corresponding covariances. It does not make the inputs independent, normally distributed or interchangeable. The variables still represent three distinct quantities whose joint behaviour needs a defensible account.
Write a = Corr(X,Y), b = Corr(X,Z) and c = Corr(Y,Z). The important word is jointly: the three numbers should describe one set of quantities under the same conditions. Correlations from incompatible populations or different uncertainty definitions cannot automatically be assembled into one budget.
The first learning checkpoint is therefore conceptual. Identify what each input represents and what a repeated realisation of the complete three-input system would mean. A matrix with labels but no common observation or uncertainty model can conceal a problem before any arithmetic begins. The equations will test coherence, but they cannot supply the missing meaning of the inputs.
The correlation matrix R has rows [1, a, b], [a, 1, c] and [b, c, 1]. The diagonal entries are one because each standardised input has unit variance. The repeated entries express symmetry: the correlation of X with Y equals that of Y with X.
A pairwise check requires every off-diagonal entry to lie between minus one and one. That condition is necessary. It is sufficient when only two standardised variables are involved, but three variables create a further compatibility requirement.
Imagine saying that X and Y move almost identically, and that X and Z also move almost identically. It becomes difficult to maintain that Y and Z usually move in strongly opposite directions. The matrix condition gives this intuition an exact form, while avoiding the assumption that a verbal picture alone proves a numerical bound.
Keep the row order visible. Swapping the labels on just one column changes the table rather than relabelling it consistently. If the inputs are reordered, rows and columns must move together. This practical check is simple but useful when uncertainty budgets are combined from several files or methods.
CHAPTER 3 OF 26 · Build the joint constraint
3. Derive the non-negative variance requirement
Take any fixed real weights h₁, h₂ and h₃ and form L = h₁X + h₂Y + h₃Z. Its variance is the expected square of its centred value, so it cannot be negative. Expanding that square gives Var(L) = hᵀRh.
A symmetric matrix is positive semidefinite when hᵀRh is non-negative for every real vector h. This is why the condition belongs to covariance and correlation models: it expresses a property every possible weighted output must have.
The word every matters. Checking one favoured output is insufficient. An inconsistent matrix can give a positive variance for the sum while giving a negative variance for a different combination. The contradiction may remain hidden until a later user asks another question of the same inputs.
This argument needs finite second moments, rather than a normal distribution. Normality becomes relevant for other interpretations, such as a complete Gaussian joint model or particular interval procedures. The basic non-negative variance check applies more broadly. A calculation should state which conclusions use only second moments and which require stronger distributional assumptions.
CHAPTER 4 OF 26 · Build the joint constraint
4. Construct a table that passes every pairwise check
Consider the fictional proposal a = 0.9, b = 0.9 and c = −0.9. Every entry is within the ordinary correlation range. Each separate two-input correlation matrix is therefore admissible by itself.
The proposal says that Y and Z each track X strongly in the positive direction, while tracking one another strongly in the negative direction. It feels suspicious, but the useful next move is a reproducible contradiction.
Choose L = −X + Y + Z. The three squared-weight terms contribute three. The cross terms contribute 2(−a − b + c). Substituting gives Var(L) = 3 + 2(−0.9 − 0.9 − 0.9) = −2.4.
No real random variable with finite variance can have variance minus 2.4. Consequently the three proposed correlations cannot all hold in one joint model. This conclusion does not identify which estimate is wrong. It identifies a conflict among them. Choosing the most convenient entry to discard would be an additional decision requiring evidence about how the entries were obtained and which assumptions deserve revision.
CHAPTER 5 OF 26 · Build the joint constraint
5. Check why one positive output is not enough
Using that same invalid table, the simple sum X + Y + Z has computed variance 3 + 2(a + b + c) = 4.8. A user interested only in the sum might see a positive answer and assume the budget is coherent.
The difference Y − Z also gives a positive number: 2 − 2c = 3.8. Neither calculation repairs the negative variance of −X + Y + Z. A table cannot become a valid joint covariance model just because the current output avoids its contradictory direction.
This distinction is practically useful in reusable modelling. A dependence budget may serve several downstream calculations, some involving additions and others involving differences. Its validity should be checked at the input-model level, rather than rediscovered separately whenever an output happens to fail.
Ask a learner to explain why the positive 4.8 does not cancel the negative −2.4. They are variances of different proposed combinations, not positive and negative contributions to one final account. Every individual variance must be non-negative. Adding the two numbers would erase the logical requirement the model was supposed to satisfy.
CHAPTER 6 OF 26 · Derive and construct valid models
6. Obtain the three-input determinant condition
For the three-input unit-diagonal matrix, the determinant is 1 + 2abc − a² − b² − c². A valid correlation matrix must have a non-negative determinant, alongside its other non-negative principal minors.
For this particular three-by-three setting, symmetry, unit diagonal, entries within the pairwise range and a non-negative determinant together are sufficient for positive semidefiniteness. This is a low-dimensional result; a determinant alone is not a general validity test for larger matrices.
In the invalid example the determinant is 1 + 2(0.9)(0.9)(−0.9) − 0.81 − 0.81 − 0.81 = −2.888. It agrees with the direct variance contradiction. The determinant provides a compact check, while the weighted-variable example explains what fails physically or statistically.
Do both once by hand. Then use an appropriate numerical method for larger budgets, retaining tolerances and the matrix’s scale. A software output saying invalid is easier to investigate when the reader understands the underlying requirement. Conversely, a software routine accepting the supplied entries does not establish that it actually tested the joint condition.
CHAPTER 7 OF 26 · Derive and construct valid models
7. Solve for the allowable third correlation
Rewrite the determinant inequality as (c − ab)² ≤ (1 − a²)(1 − b²). Taking the non-negative square root gives the interval ab − √[(1 − a²)(1 − b²)] ≤ c ≤ ab + √[(1 − a²)(1 − b²)].
With a = b = 0.9, the centre is 0.81 and the half-width is 0.19. Therefore c must lie between 0.62 and 1. The proposed value −0.9 is far outside this interval, even though it is acceptable as an isolated pairwise correlation.
The bound answers a conditional question: if the first two correlations are exactly 0.9, these are the third correlations compatible with the second-moment model. It does not establish the actual third correlation. Any empirical estimate inside the interval still needs evidence, uncertainty assessment and an appropriate population definition.
Use the formula after defining the quantities. It should not become a way to force unrelated datasets into an apparently coherent table. Coherence is a necessary modelling discipline. It is not evidence that the model’s dependence pattern describes the system under study.
CHAPTER 8 OF 26 · Derive and construct valid models
8. Explain the bound through residuals
Remove the part of Y and Z linearly aligned with X by defining E = Y − aX and F = Z − bX. Their variances are 1 − a² and 1 − b². Their covariance is c − ab.
For any two finite-variance variables, the squared covariance cannot exceed the product of their variances. Applying that condition to E and F gives (c − ab)² ≤ (1 − a²)(1 − b²), the same allowed interval as before.
This derivation makes the constraint easier to interpret. After accounting for their shared linear association with X, the two residuals have only a limited amount of variation left. Their remaining covariance cannot exceed what that residual variation permits.
For a = b = 0.9, each residual variance is 0.19. A proposal c = −0.9 would require residual covariance −1.71, far larger in magnitude than the allowed 0.19. The conflict is therefore not a vague impression about strong relationships. It is an impossible demand on the residual variation remaining after the shared component has been removed.
CHAPTER 9 OF 26 · Derive and construct valid models
9. Construct a valid interior example
Choose three independent, mean-zero, unit-variance variables U, V and W. We can construct X = U and Y = 0.9U + √0.19 V. These definitions give unit variances and Corr(X,Y) = 0.9 directly.
To obtain c = 0.8 while keeping Corr(X,Z) = 0.9, define Z = 0.9U + [(0.8 − 0.81)/√0.19]V + √[0.19 − (0.8 − 0.81)²/0.19]W. The expression inside the last square root is positive.
The V coefficient is approximately −0.022942 and the W coefficient approximately 0.435286. Expanding variances verifies Var(Z) = 1. Expanding Cov(Y,Z) gives 0.81 plus −0.01, or 0.8. This is a complete construction, rather than a table merely declared plausible.
Independent standard normal U, V and W would additionally produce a joint normal example. But normality is not needed for these second-moment checks. Other independent inputs with the stated moments also work. Keeping this distinction explicit prevents a constructive proof of covariance validity from quietly becoming a claim that the distribution’s shape is known.
CHAPTER 10 OF 26 · Derive and construct valid models
10. Understand the lower boundary
Set c = 0.62 with a = b = 0.9. A valid construction is X = U, Y = 0.9U + √0.19 V and Z = 0.9U − √0.19 V, using two independent standardised variables.
Adding Y and Z gives exactly 1.8X. Thus Y + Z − 1.8X is identically zero and has variance zero. This is a valid dependence relationship, not a negative variance or a contradiction.
The resulting correlation matrix is singular: it has a direction with no variation. Positive semidefinite permits such directions. Positive definite requires strictly positive variance for every non-zero weight vector and therefore excludes this exact linear relationship.
A numerical routine that requires a positive definite matrix may reject this valid boundary model. The message then concerns the routine’s stricter requirements or the operation being attempted, such as an ordinary inverse. It does not by itself show that the covariance description is mathematically impossible. Distinguish an invalid model from a valid model unsuitable for a particular algorithm before deciding what to change.
CHAPTER 11 OF 26 · Inspect provenance and transformations
11. Understand the upper boundary
At c = 1, Y and Z can be exactly equal as standardised variables. For example, use X = U and Y = Z = 0.9U + √0.19 V. Their correlations with X are both 0.9, and their mutual correlation is one.
The difference Y − Z has variance zero. This does not say that the two original physical quantities have identical units or raw readings. It says that their centred, standardised deviations are identical under the constructed uncertainty model.
A learner may assume that perfect correlation means X, Y and Z are all identical. The example shows otherwise. X has a separate relationship with the other two, and its correlation with them is only 0.9. Perfect dependence of one pair does not impose the same relationship on every pair.
Compare the two boundaries. At the lower one, a weighted sum of all three variables is fixed. At the upper one, a pairwise difference is fixed. Both have determinant zero but different zero-variation directions. The determinant flags singularity; the underlying variable relationship explains what the singularity means.
CHAPTER 12 OF 26 · Inspect provenance and transformations
12. Inspect a compact set of alternatives
| Proposed a and b | Proposed c | Joint second-moment status | A useful check |
|---|---|---|---|
| 0.9 and 0.9 | −0.9 | Invalid | Var(−X + Y + Z) = −2.4 |
| 0.9 and 0.9 | 0.62 | Valid, singular | Y + Z − 1.8X can be identically zero |
| 0.9 and 0.9 | 0.8 | Valid, positive definite | Construct with three independent inputs |
| 0.9 and 0.9 | 1 | Valid, singular | Y − Z can be identically zero |
Read the table as a collection of separate models. Its rows are not successive observations from one experiment, and the boundary relationships describe constructions satisfying the stipulated moments. They do not establish which model fits a real dataset.
The table also distinguishes two tasks. Validity asks whether a joint model with these moments can exist. Estimation asks which dependence values are supported by evidence. A mathematically valid guess can still be empirically poor. An empirically assembled set of pairwise estimates can be jointly invalid. Keeping both questions visible is more useful than treating one matrix check as a complete uncertainty analysis.
CHAPTER 13 OF 26 · Inspect provenance and transformations
13. See how pairwise deletion can create a conflict
Consider six fictional records with missing entries. Records one and two contain X,Y pairs (−1,−1) and (1,1). Records three and four contain X,Z pairs (−1,−1) and (1,1). Records five and six contain Y,Z pairs (−1,1) and (1,−1). The third variable is missing in every record.
Computing each correlation from its own available two-record subset gives a = 1, b = 1 and c = −1. Every pairwise calculation is arithmetically correct for its particular subset. Their assembly is not a possible joint correlation model.
If Y equals X and Z equals X in one joint standardised model, Y and Z cannot also be opposites. The separate records do not supply one complete shared observation basis. Their tiny sizes make them poor estimates too, but the coherence issue would remain conceptually distinct even with larger incompatible subsets.
This example does not prove that missing data must always be discarded. It shows why a missingness strategy is part of the model. Appropriate estimation can use incomplete data while respecting joint constraints. Simply filling one matrix with independently computed pairwise entries does not automatically achieve that.
CHAPTER 14 OF 26 · Inspect provenance and transformations
14. Distinguish a common-data covariance estimate
When a sample covariance matrix is computed consistently from the same complete rows, using one common non-negative weighting scheme and a suitable common centring convention, its quadratic form can be expressed as a non-negative sum of squared weighted deviations.
That structure explains why such an estimate is positive semidefinite in exact arithmetic. Its validity does not mean it estimates the population covariance accurately, nor that the observations are independent. It means the reported second moments are collectively possible for the data representation being used.
Pairwise deletion can break the common squared-deviation representation because different entries use different rows, centres and denominators. Mixing an old calibration covariance with a new population correlation can create another compatibility problem. Trace the provenance of the entries before interpreting an invalid matrix as a mysterious software failure.
A practical record therefore identifies the observation set, weighting, centring, missingness treatment and transformation used for each component. Where the components come from different sources, a joint modelling argument is needed. A longer table of source names is useful only when it makes those relationships inspectable.
CHAPTER 15 OF 26 · Inspect provenance and transformations
15. Return from correlations to physical uncertainty
Suppose the physical input standard uncertainties are 0.20, 0.30 and 0.40 in their respective units. Let D be the diagonal matrix containing those positive values. The covariance matrix is Σ = DRD.
For any physical output weights g, its variance is gᵀΣg = (Dg)ᵀR(Dg). Thus a valid correlation matrix gives a valid covariance matrix after this rescaling. Because D is invertible here, rescaling cannot turn an invalid correlation matrix into a valid one.
The standardised contradictory direction can be recovered by using physical weights proportional to −1/0.20, 1/0.30 and 1/0.40. In a genuine physical model, coefficients must also carry the units needed to express a common output. Writing a formal weighted combination is not permission to add incompatible raw quantities without that relationship.
This matters when a budget looks reassuring because every uncertainty is small. Magnitudes can change the numerical size of a contradiction; they do not remove its sign. Coherence concerns the complete relationship, while the units and sensitivities determine how that relationship affects a particular output.
CHAPTER 16 OF 26 · Inspect provenance and transformations
16. Change a sign convention consistently
Suppose a sensor’s direction convention is reversed, so the new variable is Y* = −Y. Its correlations with X and Z both change sign. The X,Z correlation remains unchanged, and the diagonal variances remain one.
In matrix form, multiply both the relevant row and column by minus one. This is a consistent transformation of the variables and preserves positive semidefiniteness. Changing just one entry because a plotted line points downward is not the same operation.
Apply the check to a valid model a = b = 0.9, c = 0.8. After reversing Y, the correlations become −0.9, 0.9 and −0.8. The transformed model remains valid even though two signs have changed. Validity does not require correlations to be positive.
For a practical audit, record whether a coefficient describes the input’s physical orientation or the output model’s sensitivity. A negative output sensitivity already changes how covariance contributes to the result; it should not also be used to reverse a supplied input correlation unless the input variable itself has been redefined. Otherwise the same sign change can be applied twice.
CHAPTER 17 OF 26 · Test model choices and limits
17. Interpret eigenvalues without losing the meaning
A symmetric matrix can be examined through its eigenvalues. Positive semidefiniteness requires all of them to be non-negative, with an appropriate numerical tolerance in computation. A negative eigenvalue identifies a direction assigned negative variance.
Our invalid matrix with a = b = 0.9 and c = −0.9 has eigenvalues −0.8, 1.9 and 1.9. The vector (−1,1,1) lies in the negative direction; its squared length is three, giving the earlier quadratic value three times −0.8, or −2.4.
A very small negative computed value needs context. It may reflect floating-point error around a theoretically zero direction, rounding of reported entries or a substantive incompatible model. The appropriate tolerance depends on scale, numerical method and intended use. It should not be chosen simply to make a disputed matrix pass.
An eigenvalue report is therefore evidence for investigation, not a substitute for provenance. Keep both the numerical result and the input construction. A model with a large contradiction cannot be explained away as round-off, while a singular exact model should not be condemned solely because finite precision prints a tiny negative number.
CHAPTER 18 OF 26 · Test model choices and limits
18. Treat numerical repair as a changed model
One possible alteration is Rλ = λR + (1 − λ)I, shrinking off-diagonal entries towards zero while retaining unit diagonal. This operation changes the proposed dependence. It is a model revision, even when a program applies it automatically.
For our invalid example, the smallest eigenvalue becomes 1 − 1.8λ. It is non-negative when λ is at most 5/9. At λ = 5/9, the off-diagonal values become 0.5, 0.5 and −0.5, forming a valid singular matrix.
That calculation gives a transparent mathematical boundary for this chosen revision family. It does not prove that the revised correlations are the best estimates. A different loss function, reliability weighting or scientific model can justify a different alteration.
Nicholas Higham’s research on the nearest correlation matrix supplies primary context for constrained numerical approximation. The distinction to preserve is simple: closeness to the original table under a specified norm is a computational objective. Agreement with the real joint uncertainty requires evidence beyond that objective. Do not conceal an altered dependence model inside a successful software message.
CHAPTER 19 OF 26 · Test model choices and limits
19. Preserve uncertainty about the correlations themselves
Suppose a and b are fixed at 0.9, while c is described only as lying between 0.5 and 0.8. The joint-validity constraint reduces the admissible c range to 0.62 through 0.8. The original lower part is incompatible with the fixed first two values.
For the output Y − Z, variance is 2 − 2c. Across the reduced range it runs from 0.4 to 0.76. These are conditional bounds on a variance, not a probability distribution over variances. A uniform draw of c over that interval would introduce an additional assumption.
If a and b are uncertain too, every candidate combination must satisfy the joint condition. Drawing all three independently from separate intervals can generate invalid combinations even when each interval contains plausible values. Dependence among the parameter choices is part of the admissible model set.
This is a useful transfer from input uncertainty to uncertainty about the input model. More decimal places in a correlation estimate do not remove sampling uncertainty or incompatible provenance. State what is fixed, what varies and which constraints connect those choices before interpreting a sensitivity envelope.
CHAPTER 20 OF 26 · Test model choices and limits
20. Solve a new correlation-range problem
For an independent exercise, set a = 0.6 and b = 0.8. Before calculating, ask whether c = 0 is possible. Both variables are positively associated with X, but the amount of remaining variation differs from the earlier 0.9 example.
The centre ab is 0.48. The half-width is √[(1 − 0.36)(1 − 0.64)] = √(0.64 × 0.36) = 0.48. The allowed interval is therefore zero through 0.96. In this case zero is a valid boundary value.
A construction at c = 0 uses X = U, Y = 0.6U + 0.8V and Z = 0.8U − 0.6V. Each variance equals one, and Cov(Y,Z) = 0.48 − 0.48 = 0. The linear relationship 0.6Y + 0.8Z = X reveals the singular direction.
Now test c = −0.1. It lies within the ordinary pairwise range but outside the conditional allowed interval, so the complete model is invalid. The result demonstrates why a memorised rule such as positive correlations require every correlation to be positive is too crude. Derive the actual range for the supplied values.
CHAPTER 21 OF 26 · Test model choices and limits
21. Build a practical checking sequence
Start with the common question and input definitions. Check that variances are finite and that every correlation corresponds to the same stated quantities. Verify symmetry, diagonal entries, labels and pairwise limits before testing the joint condition.
For three inputs, a hand-derived allowed interval provides a useful independent check. For larger models, use an appropriate symmetric-matrix method and retain numerical tolerances. Investigate substantial negative directions through data provenance and model construction rather than merely replacing the final negative output variance with zero.
If a revision is made, save the original and revised matrices, the reason for the change and the effects on relevant outputs. A repair that barely changes one output can substantially change another. The model may be reused later by someone asking a different question.
For learners, begin with the explicit −2.4 contradiction. Stabilise understanding by constructing a valid boundary model. Then extend to rescaling, changed signs and uncertain correlations. The aim is independent method selection: recognise when a table needs a joint check and explain why the chosen operation addresses that issue. Correct arithmetic becomes more durable when its underlying possibility condition is visible.
CHAPTER 22 OF 26 · Practise and continue
22. Handle a perfect first correlation without dividing by zero
The residual formula sometimes tempts a learner to divide by √(1 − a²). That step is unavailable when a equals one or minus one. The original covariance relationship still has meaning; the derivation must respect its boundary.
If a = 1, the centred standardised variables Y and X are equal almost surely. This follows because Var(Y − X) = 2 − 2a = 0. Therefore Corr(Y,Z) must equal Corr(X,Z), so c = b. The allowed-range formula gives the same result: its half-width is zero and its centre is b.
For a fictional example, set a = 1 and b = 0.3. The only compatible third correlation is c = 0.3. Choosing c = 0.4 does not become acceptable because the discrepancy is only one tenth; it violates the stipulated exact relationship. An uncertain estimate near one is a different situation and should not be silently treated as exact perfection.
If a = −1, Y = −X almost surely in this standardised model, giving c = −b. These cases teach a broader mathematical habit: a formula containing division has a domain. When its denominator vanishes, return to the underlying relationship instead of performing the forbidden operation or replacing the zero with an arbitrary tiny number. A boundary often has a simple interpretation that the unstable algebra would obscure.
CHAPTER 23 OF 26 · Practise and continue
23. Distinguish a valid pooled model from the current regime
Suppose two fictional operating regimes each have standardised inputs with the same zero means and unit variances. Regime A has identity correlation matrix: its pairwise covariances are zero. Regime B has all three pairwise correlations equal to 0.8. Both matrices are valid.
If a record selects either regime with equal probability and preserves those common means and variances, the pooled correlation matrix has off-diagonal entries 0.4. It is a valid joint model for that mixture. Its validity does not mean that 0.4 describes the inputs conditional on either individual regime.
For L = X + Y + Z, the variance is three in A and 7.8 in B. The pooled value is 5.4, their equally weighted average in this particular common-mean construction. Reporting 5.4 as the uncertainty for a system known to be in B would answer the pooled question rather than the stated current question.
The common means matter. If regimes also have different means, the covariance of those changing conditional means adds another contribution. Averaging within-regime covariance matrices alone would then omit part of the pooled variation.
This example separates coherence from relevance. Every matrix in the exercise is mathematically possible, yet using the wrong one can still misstate uncertainty. The practical record should identify operating conditions, observation windows and intended population. Passing a positive-semidefinite check is the start of a usable joint model, rather than evidence that every later question should use the same pooled dependence.
CHAPTER 24 OF 26 · Practise and continue
24. Notice when two supplied correlations leave the third unresolved
Set a = b = 0. The allowed interval for c is minus one through one. The fact that X is uncorrelated with both Y and Z therefore imposes no further restriction on their pairwise correlation in this example.
Two explicit constructions make the range visible. With independent standardised U and V, use X = U, Y = V and Z = V for c = 1. Alternatively set Z = −V for c = −1. Both constructions preserve the same two zero correlations involving X.
Yet Var(Y + Z) is four in the first construction and zero in the second. The two given correlations are insufficient to determine that output variance. Calling the unspecified third correlation zero would insert an assumption, rather than complete a calculation from the supplied evidence.
This example complements the earlier incompatibility problem. Sometimes a proposed entry is impossible; sometimes several entries are possible and consequentially different. A useful uncertainty analysis distinguishes those states. The next investigation should seek dependence information about the unresolved pair, instead of collecting more evidence only about the two relationships already known.
CHAPTER 25 OF 26 · Practise and continue
25. Questions readers should be able to answer
Is every table of pairwise correlations a valid correlation matrix?
No. Pairwise limits are necessary, but the complete table must be positive semidefinite. The 0.9, 0.9, −0.9 example gives a direct contradiction.
Does a zero determinant mean the model is impossible?
No. A valid singular model can contain an exact linear relationship. Distinguish zero variance in one direction from negative variance.
Must the inputs be normally distributed for this check?
No. The non-negative variance requirement uses finite second moments. A full Gaussian interpretation is an additional assumption.
Does numerical projection discover the true correlations?
No. It finds a mathematically admissible revision under a specified objective. The empirical adequacy of that revision remains a separate question.
Can I ignore invalidity if my present output variance is positive?
That does not establish a coherent reusable joint model. Another weighted combination may expose the contradiction, as the worked example shows.
CHAPTER 26 OF 26 · Practise and continue
26. Continue through the connected evidence
Build the uncertainty model
Return to Uncertainty Quantification for targets, budgets, propagation and interpretation. The existing ratio confidence-set guide develops a different issue: what an admissible statistical model can imply for a nonlinear parameter ratio.
Strengthen the covariance foundation
The covariance explanation develops expected products, matrix geometry, missingness and propagation. Use it for the wider foundation; the present article supplies an explicit three-input compatibility calculation and constructive alternatives. Numerical Analysis helps distinguish approximation and conditioning from the meaning of the model.
Read the primary methodological context
Higham's July 2002 paper record describes nearest-correlation approximation. Source route checked on 23 October 2026. Our numerical matrices, residual constructions and transfer problems are original calculations rather than examples taken from that record.
Continue through the Gold Standard education parent or How X Works. Choose three uncertain inputs, explain their common observation basis and test whether their proposed relationships can all hold at once.
Contents · Previous chapter · Return to Uncertainty Quantification